The Fifth Elephant 2020 edition

On data governance, engineering for data privacy and data science

Detecting & Addressing Out of Distribution Data (OOD) Issues in Production ML Systems

Submitted by Saravanan Chidambaram (@sarochida) on Mar 25, 2020

Status: Submitted

Abstract

Deep learning systems have achieved enormous progress over the past decade in analysing and predicting text, tabular and image data. However during deployment of these systems, there has been issues in handling out of distribution (OOD) data. Deep neural networks can end up making highly confident wrong predictions when real world input data is from a distribution different from that of the training data. Such highly confident wrong predictions can impact the safety of AI applications adversely in real world deployment. There has been considerable research in (a) detecting out of distribution data (b) predicting the performance drop data under OOD condition and (c) mitigating and handling OOD data. In this talk, we discuss the current state of art methods for detecting OOD data, and cover techniques for addressing the same.

Outline

We start by discussing domain/co-variate shift and label shift concepts and point out the basic tenet of ML systems (IID principle) which gets violated with OOD data. We point out with real world examples, how ML systems fail silently with OOD inputs leading to AI safety issues. We then discuss methods for detecting dataset shift, identifying exemplars that most typify the shift, as well as quantifying the adverse impacts of the shift on system performance. We also briefly cover the work around predicting performance drop under domain shift. We then discuss monitoring ML systems in production using “data unit tests” to handle OOD issues. We briefly cover automated data quality monitoring and distributional shift detection for ML pipelines in deployment.

References (that will be covered in this talk)
https://www.irt-systemx.fr/wp-content/uploads/2020/01/19_Cl%C3%A9ment-FEUTRY.pdf
https://arxiv.org/abs/1908.04388
http://papers.nips.cc/paper/8420-failing-loudly-an-empirical-study-of-methods-for-detecting-dataset-shift
https://arxiv.org/abs/1912.05651
https://europe.naverlabs.com/research/publications/to-annotate-or-not-predicting-performance-drop-under-domain-shift/
https://ssc.io/pdf/autoops.pdf

Speaker bio

Saro is a hands-on technologist & management leader, and has two decades of experience in building enterprise software. Currently CEO of a pre-seed NLP startup (which won the Y-Combinator StartupSchool Grant of $15K). Prior to that, Saro was the Head of the Advanced Development Center, Hewlett Packard Enterprise India – part of HPE’s Global CTO office where he managed the core research team in AI and Edge Computing, collaborating with HP labs and business units to take research to products. He has led, managed, mentored high performance research and development teams in HPE and have demonstrated significant business impact of research. He has an M.Tech. degree in Computer Engineering from IIT Kharagpur.

Links

Comments

Login to leave a comment